DataClassifyGrading

A leaf-classification framework with strict JSON model outputs. The formal shougang RLOO path uses a native Qwen3.5 tool trajectory rather than a synthetic two-stage user bridge:

1. The model may answer directly or call deterministic search_categories.
2. It may optionally inspect candidates with get_category_details or get_category_examples; tool observations are excluded from policy loss.
3. The terminal assistant response is exactly {"answer":"<choice_id>", "level":"<data_level>"} and receives one strict joint exact-match reward.

The tools read only the runtime registry/corpus JSON. They use no embedding model, vector store, network retrieval, or per-sample ground truth.

Repository data policy

The repository contains code, engineering configuration, synthetic examples/tests, and the approved data/knowledge/**/*.json knowledge base. Production datasets, registries, corpora, model inputs, reports, and training artifacts must remain local and untracked.

Fresh-clone verification

Install the test extra so a fresh checkout has both pytest and pyarrow:

python -m pip install -e ".[test]"
python -m pytest

The examples in cfg/task/*.example.json are fabricated and are suitable for smoke tests only. The VeRL server environment is separate; the native-tool RLOO launcher requires the validated core in requirements/verl-qwen35-cu130.txt (VeRL 0.9.0, vLLM 0.27.1, Transformers 5.10.4, PyTorch 2.13.0+cu130) on the authorized RTX PRO 6000 96GB host. Do not mix it with the archived VeRL 0.8 environment.

Local classification

Supply every production asset by explicit path. Formal shougang runs use a task config whose metadata is field_name and table_name, plus a grading config whose ground-truth field is data_level:

python -m script.verl.sft.evaluate_true_e2e \
  --model-path <local-model-dir> \
  --data <local-test.parquet> \
  --registry <local-registry.json> \
  --corpus <local-corpus.json> \
  --task-config <local-task.json> \
  --grading-config <local-grading.json> \
  --report <local-output-report.json>

--report is required and the output is runtime-local. Nothing under runtime paths (datasets, standards, registries, corpora, parquet, models, reports, or checkpoints) should be added to Git. The reference example uses /absolute/path/to/Qwen3.5-9B; old Qwen2.5 token limits are invalid, so re-measure the selected model's chat-template budget with the repository token budget CLI before training. Start with docs/classification-runtime.md and follow the exact server sequence in docs/server-runbook.md.